PGSMamba: Prompt-Guided Shuffle State Space Model for Hyperspectral Image Classification

Hyperspectral image classification demands models capable of efficiently capturing complex spectral–spatial relationships and long-range dependencies. Despite significant advances in CNNs and Transformers, balancing modeling capacity with computational efficiency remains challenging. Recently, Mamba has emerged as a promising alternative with linear-complexity global modeling capabilities. However, current Mamba-based methods suffer from two critical limitations: artifact tokens that compromise foreground-background distinction, and overfitting tendencies exacerbated by their non-hierarchical, homogeneous layer architecture. To address these limitations, we propose PGSMamba, a prompt-guided shuffle state space model for hyperspectral image classification. Our approach incorporates three core components: a patch-prompt sequence construction module for artifact filtering through learnable prompts, a shuffle scanning module providing layer- wise regularization via random perturbations, and a prompt-to-patch aggregation strategy enabling robust global-local feature fusion and effectively suppressing artifact tokens within patch tokens. Comprehensive experiments on five benchmark datasets, including Indian Pines, Pavia University, Houston 2013, WHU-Hi-LongKou, and Salinas, demonstrate that PGSMamba achieves state-of-the-art overall accuracies of 97.88%, 98.09%, 94.02%, 98.92%, and 95.11%, respectively. These results confirm the effectiveness and generalization ability of our approach, highlighting PGSMamba as a promising solution for efficient and robust hyperspectral image classification.